This paper presents a stock prediction method that integrates news feature extraction with recurrent neural networks (RNNs). The authors first select seed words indicative of good or bad news and optimize word polarity scores. News features are then constructed based on the aggregated polarity of words in each article. To model the sequential nature of stock prices and the continuous impact of news, an RNN is employed. Experiments show that the proposed method achieves over 5% higher accuracy compared to a standard SVM classifier using only price features, demonstrating the value of incorporating news sentiment into time-series prediction.